MBD-Net: Multi-Branch Dilated Convolutional Network With Cyst Discriminator for Renal Multi-Structure Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate three-dimensional (3D) kidney parsing from computed tomography angiography (CTA) is crucial for surgery-based renal cancer treatment.
• Existing methods may face challenges in precisely segmenting complex renal structures like tumors and cysts.

Purpose of the Study:

• To propose an end-to-end deep learning framework for automated segmentation of multiple renal structures from CTA images.
• To develop a novel network architecture capable of efficient and accurate feature extraction for renal structure identification.
• To create a discriminator module that differentiates renal tumors from cysts without requiring manual labels.

Main Methods:

• An encoding-decoding convolutional neural network (CNN) framework, termed Multi-Branch Dilated Convolutional Network (MBD-Net), was developed.
• MBD-Net incorporates residual, hybrid dilated convolutional, and reduced-dimensional convolutional structures for enhanced feature extraction with fewer parameters.
• A Cyst Discriminator module was designed to distinguish tumors from cysts using grayscale curves and radiographic features.

Main Results:

• The framework achieved high segmentation accuracy on the MICCAI 2022 KiPA2022 dataset.
• Mean Dice Similarity Coefficients (DSC) were 96.18% for kidneys, 90.99% for kidney tumors, 88.66% for arteries, and 80.35% for veins.
• The approach demonstrated stable and top performance in the challenge.

Conclusions:

• The proposed CNN-based framework effectively automates the segmentation of 3D kidneys, tumors, arteries, and veins from CTA images.
• This automated segmentation aids in kidney parsing, offering significant benefits for surgery-based renal cancer treatment.
• The method shows promise for improving pre-operative planning and surgical outcomes in oncology.